Roadmap
This page summarizes the repository's ROADMAP.md, which is authoritative. Everything on this page that is not marked implemented is planned and does not work today.
Release status
| Version | Theme | Status |
|---|---|---|
| 0.1–0.2 | Registry, discovery, HTTP API, CLI, status, snapshots, SQLite, tokens | Implemented |
| 0.3 | Read-only AgenticOps Console | Implemented |
| 0.4 | Controlled configuration and approvals | Implemented |
| 0.5 | Ecosystem evidence ingestion and posture | Implemented |
| 0.6 | MCP investigation and approved staging recovery | Implemented |
| 0.7 | Multi-agent operations and bulk actions | Implemented |
| 0.8 | Live detection, alerts, audit trails and incident views | Implemented (0.8.0 on PyPI) |
| 0.9 | Automated runbooks and recovery workflows | Planned |
| 1.0 | Stable control plane and published capability contract | Planned |
0.9: automated runbooks (planned)
- Versioned runbooks with prerequisites, scoped targets, ordered actions, human approval gates, bounded retries and timeouts, and rollback/reconciliation.
- The copilot recommends supported runbooks without inventing commands or extending an operator's authority.
- Execution and verification after approval, with cancellation, interruption recovery and escalation.
- Runbooks validated against controlled Chaos scenarios, with Evals results recorded.
1.0: stable contract (planned)
- A published, stable capability-discovery contract and locked core API semantics.
- Documented runtime and framework integration patterns, with at least two materially different runtimes validated end to end.
Other planned work
From the repository's operational limits:
- Timeout-based health and automatic package detection.
- Pull-based evidence collection and remote configuration application.
- Production runtime adapters and production runbooks.
- Distributed storage, per-user identities and tenant isolation.
- Per-user credential binding so the hosted MCP endpoint can offer Tower tools.
Vision
Detect agent failures → investigate live evidence through MCP → explain evidence-backed causes → get human approval → run a bounded runbook → verify recovery. 0.6–0.8 deliver the first loop for one staging runtime; 0.9 makes it reusable.